A parameter estimator based on Smoluchowski–Kramers approximation

A parameter estimator based on Smoluchowski–Kramers approximation

Accepted Manuscript A parameter estimator based on Smoluchowski-Kramers approximation Ziying He, Jinqiao Duan, Xiujun Cheng PII: DOI: Reference: S0...

468KB Sizes 0 Downloads 85 Views

Accepted Manuscript A parameter estimator based on Smoluchowski-Kramers approximation

Ziying He, Jinqiao Duan, Xiujun Cheng

PII: DOI: Reference:

S0893-9659(18)30351-3 https://doi.org/10.1016/j.aml.2018.10.011 AML 5669

To appear in:

Applied Mathematics Letters

Received date : 27 September 2018 Revised date : 18 October 2018 Accepted date : 18 October 2018 Please cite this article as: Z. He, et al., A parameter estimator based on Smoluchowski-Kramers approximation, Appl. Math. Lett. (2018), https://doi.org/10.1016/j.aml.2018.10.011 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

*Manuscript Click here to view linked References

A parameter estimator based on Smoluchowski-Kramers approximationI Ziying Hea , Jinqiao Duanb,1 , Xiujun Chenga a

Center for Mathematical Sciences & School of Mathematics and Statistics & Hubei Key Laboratory for Engineering Modeling and Scientific Computing Huazhong University of Sciences and Technology, Wuhan 430074, China b Department of Applied Mathematics, Illinois Institute of Technology, Chicago, IL 60616, USA

Abstract We devise a simplified parameter estimator for a second order stochastic differential equation by a first order system based on the Smoluchowski-Kramers approximation. We establish the consistency of the estimator by using Γ-convergence theory. We further illustrate our estimation method by an experimentally studied movement model of a colloidal particle immersed in water under conservative force and constant diffusion. Keywords: Parameter estimation, second order stochastic differential equation, Smoluchowski-Kramers approximation, Γ-convergence.

1

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

1. Introduction We estimate an unknown system parameter in the drift term of a stochastic system satisfying Newton’s law by Smoluchowski-Kramers approximation. The Smoluchowski-Kramers approximation was introduced by Kramers [7] and Smoluchowski [12] to compute a small parameter limit for irregularly moving particles. It has been extended to more general systems [3]. The Smoluchowski-Kramers approximation is the main justification for replacement of the second order movement equation by the first order force-distance relation equation in physical experiments. It has been used to devise a filtering method to cut the dimension by half and bypass the dependence on velocity observation [10]. In this Letter, we devise a parameter estimator for a second order stochastic system by its first order Smoluchowski-Kramers approximation system. We verify the consistency property of the estimator by employing Γ-convergence theory, which was developed to study the optimization problems [8]. This Γ-convergence theory is about a family of functionals’ convergence to a limit functional, and the corresponding convergence for the minimizers. In general, this latter convergence can not be assured. However, in our case, the objective function is verified to be Γ-convergent, and the corresponding minimizer converges to the true parameter value. Thus the consistency about the estimator holds. We apply our parameter estimation method to determine the force in an experimental system. The force can be measured by modeling the particle with small mass through the force-distance relationship [13]. I

This work was partly supported by the NSF grant 1620449 and the NSFC grants 11531006 and 11771449. Email addresses: [email protected] (Ziying He), [email protected] (Jinqiao Duan), [email protected] (Xiujun Cheng) 1 Corresponding author ([email protected])

Preprint submitted to Elsevier

October 18, 2018

19

20 21

2. Settings and assumptions We consider the parameter estimation on ϑ in the parameter space Θ ⊂ R in the following second order Newton equation of motion under random fluctuations ˙ t, µ¨ xµ (t) = b(xµ (t), ϑ) − γ x˙ µ (t) + σ W

22 23 24 25

26 27

1 σ ˙ b(x(t), ϑ) + W t, γ γ

x(0) = x0 ∈ Rn .

31 32 33 34

a.s.,

(2.4)

for all v0 ∈ Rn , uniformly for t in compact subintervals of [0, ∞). Suppose we have available the observations {xµ (tk )}nk=1 on time interval [0, T ] with partition 1 ∆n = {0 = t1 < · · · < tk < · · · < tn = T }, ∆k t = tk+1 − tk (k = 1, 2, · · · , n − 1) and T = ∆n 2 with a positive constant ∆ [6]. We can estimate ϑ by the first order n-dimensional system (2.3). We devise a least square estimator for ϑ in equation (2.1) by equation (2.3). Define the objective function Fnµ (ϑ) =

n X ||xµ (tk ) − xµ (tk−1 ) − γ1 b(xµ (tk−1 ), ϑ)∆k t||2

∆k t

k=1

35

(2.3)

By revising the proof in [9] (page 59), we have lim xµ (t) = x(t)

30

(2.1)

The Smoluchowski-Kramers approximation assures that equation (2.1) converges in some sense to the following Smoluchowski equation [3], as µ → 0,

µ→0 29

x˙ µ (0) = v0 ∈ Rn ,

where Wt is an n-dimensional Brownian motion [2] defined on sample space Ω with probability P , γ is the friction coefficient (a positive constant), σ is a constant, and b is a continuous drift function from Rn × Θ to Rn . The second order equation (2.1) can be rewrite as a first order system  µ x˙ (t) = v µ (t), xµ (0) = x0 ∈ Rn , (2.2) ˙ t, v µ (0) = v0 ∈ Rn . v˙ µ (t) = µ1 b(xµ (t), ϑ) − µγ v µ (t) + σµ W

x(t) ˙ = 28

xµ (0) = x0 ∈ Rn ,

.

(2.5)

Then, as µ → 0, this objective function will tend almost surely to the following function Fn (ϑ) =

n X ||x(tk ) − x(tk−1 ) − γ1 b(x(tk−1 ), ϑ)∆k t||2

∆k t

k=1

.

(2.6)

This is the objective function of the least square estimation for equation (2.3) with observations {x(tk )}nk=1 . We denote the true parameter value as ϑ0 , and the unique minimizer for Fn (ϑ) as ϑn . Then under suitable conditions, this estimator possesses the consistency property [6] ϑn → ϑ0

a.s.

as n → ∞.

Let ϑµn minimize Fnµ (ϑ). We want to infer the consistency property ϑµn → ϑ0

a.s.

as µ → 0, n → ∞.

36

To this end, we make the following assumptions as in [6]:

37

A1. The parameter space Θ is a compact subspace of R. 2

A2. The drift term b(x, ϑ) is Lipschitz in both variables, that is, there exist continuous functions K(·) and C(·) such that ||b(x, ϑ) − b(y, ϑ)||Rn ≤ |K(ϑ)| · ||x − y||Rn , ||b(x, ϑ) − b(x, ϕ)||Rn ≤ |C(x)| · |ϑ − ϕ|,

for all ϑ, ϕ ∈ Θ, x, v ∈ Rn . Here

sup |K(ϑ)| = K < ∞, and E|C(x0 )|m = Cm < ∞

for some m > 16,

ϑ∈Θ

with E the expectation with respect to P. The following growth condition is also satisfied for some constant L. ||b(x, ϑ)||2Rn ≤ L(1 + ||x||2Rn ). A3. The process xt = (xt , t ≥ 0) is stationary and ergodic with E||x0 ||m Rn < ∞

for some m > 16.

A4. Recall that ϑ0 is the true parameter value in (2.1). Then E||b(x0 , ϑ) − b(x0 , ϑ0 )||2Rn = 0 38

39 40

41 42

43 44 45

46

47 48 49

iff

θ = θ0.

We need the following definitions to prove the consistency [8]. Definition 2.1. We say that a subset K of a topological space X is countably compact if every sequence in K has at least a cluster point in K. Definition 2.2. We say that a function is coercive on topological space X, if the closure of the set {F ≤ a} is countably compact in X for every a ∈ R. Definition 2.3. We say that the sequence (Fh ) is equi-coercive for h ∈ N (the natural number set) on topological space X, if for every a ∈ R there exists a closed countably compact subset Ka of X such that {Fh ≤ a} ⊂ Ka for every h ∈ N. 3. Consistency We first prove the Γ-convergence for the objective function Fnµk (ϑ) with respect to every sequence (µk ) tending to 0. We denote the extended real line as R = R ∪ {±∞}. Recall the definition about Γ-convergence [8]. Definition 3.1. The Γ-lower limit and the Γ-upper limit of the sequence (Fh ) are the functions from X into R defined by (Γ- lim inf Fh )(x) = sup lim inf inf Fh (y), h→∞

U ∈N (x) h→∞ y∈U

(Γ- lim sup Fh )(x) = sup lim sup inf Fh (y), h→∞

50 51 52 53

U ∈N (x) h→∞ y∈U

where U is an open subset in X and N (x) denotes the set of all open neighbourhoods of x in X. If there exists a function F : X → R such that Γ- lim inf h→∞ Fh = Γ- lim suph→∞ Fh = F , then we write F = Γ- limh→∞ Fh and we say that the sequence Fh Γ-converges to F (in X) or that F is the Γ-limit of (Fh ) (in X). 3

54

Lemma 3.2. Under the assumption A2, and for n ∈ N, Fnµ (ϑ) Γ-converges to Fn (ϑ) a.s., as µ → 0.

Proof. For every n ∈ N, the convergence (2.4) holds uniformly on [0, T ]. This implies that, for every ε > 0, there exists a positive µ0 , such that for µ ∈ (0, µ0 ), ||xµ (tk ) − x(tk )|| < ε 55

a.s.,

for k = 1, 2, . . . , n.

Hence for every n ∈ N, and the preceding µ, sup |Fnµ (ϑ) − Fn (ϑ)|

ϑ∈Θ

2   ∆k t 1 µ µ µ ||b(x (tk−1 ), ϑ) − b(x(tk−1 ), ϑ)|| ||x (tk ) − x(tk )|| + ||x (tk−1 ) − x(tk−1 )|| + ≤ sup ∆k t γ ϑ∈Θ k=1  X n ∆k t 1 (ε + ε + |K(ϑ)|ε) ≤ sup ∆k t γ ϑ∈Θ X n k=1

≤Mn ε,

for a positive constant Mn by the assumption A2. Thus for every n ∈ N, Fnµ (ϑ) → Fn (ϑ) a.s. as µ → 0, uniformly for ϑ ∈ Θ. 56 57

58 59 60 61

Note that the drift b(x, ϑ) is a continuous function of ϑ, so is Fn (ϑ). Hence Fn (ϑ) is a lower semicontinuous function on ϑ. According to Proposition 5.2 in [8], we infer the result. For the corresponding minimizers’ convergence, we still need the equi-coercive condition for the objective functions Fnµk (ϑ). Lemma 3.3. For every n ∈ N, Fnµk (ϑ) is equi-coercive with respect to k ∈ N for every sequence (µk ) satisfying µk → 0. Proof. For every ε > 0, there exists a µ0 > 0 such that, for µk ∈ (0, µ0 ), Fnµk (ϑ) > Fn (ϑ) − ε.

62 63 64 65

The function Fn (ϑ) − ε is continuous for ϑ from Θ ⊂ R to R. Therefore {Fn (ϑ) − ε ≤ a} is a closed set in space Θ for any a ∈ R, further it is also a countably compact set. That is, Fn (ϑ) is a coercive function and a lower semi-continuous function for ϑ. According to Proposition 7.7 in [8], we infer that the sequence (Fnµk (ϑ)) is equi-coercive. Theorem 3.4. Under the assumptions A1, A2, A3 and A4, the consistency property holds, i.e. ϑµn → ϑ0

a.s. as µ → 0, n → ∞.

Proof. By Corollary 7.24 in [8], together with Lemma 3.2 and 3.3, we conclude that for every n ∈ N and every sequence (µk ), ϑµnk → ϑn a.s. as µk → 0. Hence for every n ∈ N, ϑµn → ϑn

a.s. as µ → 0.

Noting the following consistency from [6], ϑn → ϑ0 66

67 68

a.s. as n → +∞,

we infer the consistency property. Hence for sufficiently large n, we can adjust the system parameter µ sufficiently small and estimate the value of ϑ0 in equation (2.1) by the observations {xµ (tk )}nk=1 and the equation (2.3). 4

69

70 71 72 73

4. An example The measurement about force is usually needed in various scientific systems [13, 5, 1, 11]. In an experimental system of a colloidal particle [13], the Smoluchowski-Kramers approximation provides a way to measure the force of the system [4]. The motion of a particle with mass µ in the presence of thermal noise is governed by the following Newton’s law [4], ( µ x˙ (t) = v µ (t), xµ (0) = 0, √ (4.1) 2kB Ttem ˙ kB Ttem µ Wt , v µ (0) = v0 . µv˙ µ (t) = F (xµ (t)) − D(x µ (t)) v (t) + √ µ D(x (t))

74 75 76 77 78 79 80 81 82 83 84

where v is the velocity, x is the position of the particle, kB is the Boltzmann constant, Ttem is the absolute temperature of the fluid, D(x) is a hydrodynamical diffusion√gradient, and γ = kB TDtem is the friction coefficient satisfing the fluctuation-dissipation relation σ = 2kB Ttem γ. The measurement for the velocity is difficult to obtain due to the irregular motion of the particle. But researchers can implement measurement about the position of trajectory of particle’s motion. For example, a single particle evanescent light scattering technique known as total internal reflection microscopy has been used [13]. The particle’s trajectory is obtained from the scattering intensities, which depend on its position relative to the interface. This measurement problem calls for a relationship between position and force instead of velocity and force. The Smoluchowski-Kramers approximation provides a theoretical basis for this [3]. Researchers can measure the force with the help of the movement of small mass particle, through the relation between position and force, x(t) ˙ =

p F (x(t))D(x(t)) ′ ˙ t, + D (x(t)) + 2D(x(t))W kB Ttem

x(0) = 0.

(4.2)

Volpe et.al [13] experimentally studied a colloidal particle immersed in water and diffusing in a closed sample cell above a planar wall, placed at x = 0. The conservative force acting on the particle is F (x, ϑ) = ϑe−κx − Gef f , 85 86 87 88 89 90

91

where ϑ is a prefactor depending on the surface charge densities, and κ−1 = 18 nm denotes the Debye length. The term Gef f = 4/3πR3 (ρp − ρs )g accounts for the effective gravitational contribution, where the diameter of the colloidal particle is 2R = 1.31 ± 0.04 µm, the density of the particle is ρp = 1.51 g/cm3 , the density of water is ρs = 1.00 g/cm3 , and the gravitational acceleration constant is g. The particle’s trajectory perpendicular to the wall is sampled with nanometer resolution. Under constant diffusion, γ and σ are constant quantities. The movement equation is depicted by system (2.2),  µ x˙ (t) = v µ (t), xµ (0) = 0, (4.3) µ µ µ ˙ t, µv˙ (t) = F (x (t), ϑ) − γv (t) + σ W v µ (0) = v0 . The mass of the particle is taken sufficiently small such that x(t) ˙ =

92 93

σ ˙ 1 F (x(t), ϑ) + W t, γ γ

x(0) = 0.

Thus the main result Theorem 3.4 holds. We use the objective function (2.5), to estimate ϑ. We make the unit of physical quantity unification as g, nm, and s. The force is x 4 1.31 2 ) · 0.51 · 9.8 · 10−3 . F (x, ϑ) = ϑ · e− 18 − π · ( 3 2

94 95

(4.4)

(4.5)

The simulation result is shown in Figure 1, which indicates that our estimator is a good approximation for the true parameter value. 5

2.5

Trajectory of the particle

×10 4

4.5

Graph of the objective function

×10 9

4 2 3.5 3

F n (ϑ)

1

2.5

µ

Position [nm]

1.5

0.5

2 1.5 1

0 0.5 -0.5

0 0

1

2

3

4

5

Time [s]

6

7

8

9

10 ×10 4

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

ϑ

Figure 1: Take γ = 61 , σ = 10, g = 9.8 m/s2 , θ0 = 0.02, n = 105 and µ = 0.001: (i) The left subfigure is one trajectory of the particle under true parameter value ϑ0 simulated through Newton motion equation (4.3). (ii) The right subfigure is the graph of the objective function (2.5) with drift force (4.5). The simulation result for the parameter ϑµn is 0.0200.

96

97 98 99

100

101 102

103 104 105

106 107 108

109 110

References [1] Brettschneider T, Volpe G., Helden L., Wehr J. and Bechinger C.: Force measurement in the presence of Brownian noise: equilibrium-distribution method versus drift method. Physical Review E, 83 (041113), 2011. [2] Duan J.: An Introduction to Stochastic Dynamics. Cambridge University Press, 2015. [3] Freidlin M.: Some remarks on the SmoluchowskiKramers approximation. Journal of Statistical Physics, 117(314):617-634, 2004. [4] Hottovy S., Volpe G. and Wehr J.: Noise-induced drift in stochastic differential equations with arbitrary friction and diffusion in the Smoluchowski-Kramers limit. Journal of Statistical Physics, 146, 762-773, 2012. [5] Hottovy S., McDaniel A., Volpe G. and Wehr J.: The Smoluchowski-Kramers limit of stochastic differential equations with arbitrary state-dependent friction. Communications in Mathematical Physics, 3 (336), 2015. [6] Kasonga R.A.: The consistency of a nonlinear least squares estimator from diffusion processes. Stochastic Processes and their Applications, 30, 263-275, 1988.

112

[7] Kramers H. A.: Brownian motion in a field of force and the diffusion model of chemical reactions. Physica, 7, 284304, 1940.

113

[8] Maso G.D.: An Introduction to Γ−Convergence. Birkh¨ auser Boston, 1993.

111

114 115

116

117 118 119

120 121

122 123

[9] Nelson E.: Dynamical Theories of Brownian Motion. Princeton University Press, second edition,1967. [10] Papanicolaou A.: Filtering for fast mean-reverting processes. Asymptotic analysis, 70, 2010. [11] Sinha P., Szilagyi I., Ruiz-Cabello F., Maroni P., and Borkovec M.: Attractive forces between charged colloidal particles induced by multivalent ions revealed by confronting aggregation and direct force measurements. The Journal of Physical Chemistry Letters, 4, 648-652, 2013. [12] Smoluchowski M.: Drei Vortr¨ age u ¨ber diffusion Brownsche Bewegung and Koagulation von Kolloidteilchen. Physik Z, 17, 557-585, 1916. [13] Volpe G., Helden L., Brettschneider T., Wehr J., and Bechinger C.: Influence of noise on force measurements. Physical Review Letters, 104 (170602), 2010. 6